Frontal face reconstruction with symmetric constraints

Frontal Face reconstruction which generates a fontal face from any pose face can produce a pose-invariant face recognition system. To this end, we propose a method for reconstructing a frontal face from a non-fontal face in a single image. Detected 2D fiducial points-based 3D Morphable Model (3DMM) fitting used for 3D face shape reconstruction. Frontal-view 3DMM is generated by rotating non-frontal-view 3DMM. The visibility of surfaces of 3DMM is measured by the corresponding visible area ratio between frontal-view and non-frontal-view 3DMM. The visible regions of frontal-view 3DMM are drew by 3D piece-wise affine warping from the face image. The remainder is how to draw invisible regions. We exploit symmetric property of face in order to draw invisible regions. Corresponding symmetric regions of invisible regions are duplicated and their global intensity is adjusted as much as the difference visible regions which have corresponding symmetric visible regions. Our proposed method is compared with a normalization method using only eye positions in terms of the performance of gender classification. Our method reaches an accuracy of 77% on NCKU database, this is 3% higher than the other method.